Full-Duplex Massive MIMO

The 150150 dB Problem

The dream is simple and has been stated in essentially the same words since 1950: a radio that transmits and receives on the same frequency at the same time doubles its spectral efficiency. Every practical radio since has been half-duplex not because of information-theoretic obstacles but because of a hardware obstacle. The transmitter, sitting inches from the receiver, is typically 101510^{15} times louder than the received signal βˆ’- about 150 dB of isolation required.

Massive MIMO changes the terms of this problem. With NtN_t transmit antennas and NrN_r receive antennas on the same array, we have a NtNrN_tN_r-dimensional self-interference MIMO channel that can be spatially nulled β€” exactly what beamforming was invented for. The question is whether the spatial nulls hold up against transmitter phase noise, nonlinear distortion, and antenna coupling variations, and how much of the 150150 dB budget they actually buy.

,

Definition:

Full-Duplex Massive MIMO System Model

A full-duplex (FD) massive MIMO base station has NtN_t transmit antennas and NrN_r receive antennas. Define the received uplink signal at the BS while the BS simultaneously transmits xDL∈CNt\mathbf{x}_{DL} \in \mathbb{C}^{N_t}:

yUL=HUL sUL+HSI xDL+HUE-UE sUE,DL+w,\mathbf{y}_{UL} = \mathbf{H}_{UL}\,\mathbf{s}_{UL} + \mathbf{H}_{SI}\,\mathbf{x}_{DL} + \mathbf{H}_{UE\text{-}UE}\,\mathbf{s}_{UE,DL} + \mathbf{w},

where:

  • HUL\mathbf{H}_{UL} is the desired uplink channel from UL users.
  • HSI∈CNrΓ—Nt\mathbf{H}_{SI} \in \mathbb{C}^{N_r \times N_t} is the self-interference (SI) channel from Tx antennas to Rx antennas on the same array.
  • HUE-UE\mathbf{H}_{UE\text{-}UE} captures cross-link interference between uplink and downlink users.

The per-antenna SI power at the receive array, before any cancellation, is typically Ptβˆ’LisolationP_t - L_{\text{isolation}}, where LisolationL_{\text{isolation}} is the passive antenna isolation (typically 3030-6060 dB for well-designed arrays) and PtP_t is the per-Tx-antenna output power.

The SI channel HSI\mathbf{H}_{SI} is deterministic and slowly varying, not random fading. It is a near-field coupling problem, not a propagation problem. This is both a blessing (it can be estimated to high accuracy) and a curse (it changes with temperature, aging, and environment, so the estimate is only ever approximately valid).

,

Definition:

The Three-Stage Self-Interference Cancellation Cascade

Commercial-grade full-duplex radios achieve their cancellation budget ΞΎ\xi via a cascaded architecture of three stages, each applied to successively cleaned residues:

  1. Passive antenna isolation (ΞΎ1β‰ˆ30\xi_1 \approx 30-6060 dB): shielding, polarization separation, physical spacing, or multi-port circulators. Removes the bulk of the SI but cannot handle multipath reflection of the Tx signal from the environment.
  2. Analog/RF cancellation (ΞΎ2β‰ˆ30\xi_2 \approx 30-5050 dB): a replica of the transmit signal, delay-matched and amplitude-weighted through an RF combiner, is subtracted from the received signal before the ADC. Crucial because the ADC's dynamic range is finite β€” without this stage, the SI saturates the ADC and destroys the received signal.
  3. Digital cancellation (ΞΎ3β‰ˆ30\xi_3 \approx 30-5050 dB): in baseband, the known transmit waveform is convolved with an estimated SI channel and subtracted from the received samples. Handles residual linear and nonlinear SI but is limited by quantization noise introduced in the ADC.

The total cancellation depth is ΞΎ=ΞΎ1+ΞΎ2+ΞΎ3\xi = \xi_1 + \xi_2 + \xi_3. For a 00 dBm received signal and 2323 dBm transmit power to become equal-SNR after cancellation with a 1010 dB noise floor, one needs ΞΎβ‰₯23βˆ’0+10=33\xi \geq 23 - 0 + 10 = 33 dB beyond the signal margin; realistic deployments target ΞΎβ‰₯110\xi \geq 110 dB, and laboratory prototypes have approached ΞΎβ‰ˆ125\xi \approx 125-130130 dB.

,

Self-Interference Cancellation Cascade

Self-Interference Cancellation Cascade
Three-stage cascade: passive antenna isolation, analog/RF cancellation, digital cancellation. Each stage operates on the residue from the previous one. Typical budgets shown for each stage; the total cancellation depth determines how much residual SI the receiver must tolerate.

Full-Duplex Self-Interference Power Budget

Trace the SI power (in dBm) through the cancellation cascade and compare to the target noise floor. Adjust the transmit power, per-stage cancellation, and noise floor to find the feasibility boundary for commercial FD massive MIMO.

Parameters
30
45
35
35
-94

Theorem: Spatial Nulling of Self-Interference (Array Gain)

Assume a FD massive MIMO BS with NtN_t Tx antennas and NrN_r Rx antennas, each arranged as a linear array with half-wavelength spacing. The SI channel HSI\mathbf{H}_{SI} has full rank min⁑(Nt,Nr)\min(N_t, N_r) and condition number κSI\kappa_{\text{SI}}. Let the downlink precoder WDL\mathbf{W}_{DL} be designed to project orthogonally to the row space of HSI\mathbf{H}_{SI}. Then the residual SI power after the spatial null is bounded by

PSI,res≀PtNtβˆ’KDLβ‹…ΞΊSI2β‹…Ο΅cal,P_{\text{SI,res}} \leq \frac{P_t}{N_t - K_{DL}} \cdot \kappa_{\text{SI}}^2 \cdot \epsilon_{\text{cal}},

where KDLK_{DL} is the number of downlink users and Ο΅cal\epsilon_{\text{cal}} is the calibration error variance of HSI\mathbf{H}_{SI}. The active antenna surplus Ntβˆ’KDLN_t - K_{DL} provides free DoF for SI suppression, giving a 10log⁑10(Ntβˆ’KDL)10\log_{10}(N_t - K_{DL}) dB cancellation improvement relative to a single-antenna system.

Massive MIMO makes self-interference a MIMO channel with a known structure. Spatial nulling uses the null space that a massive array has in abundance. The formula says that every additional Tx antenna beyond the required KDLK_{DL} streams contributes 10log⁑1010\log_{10} dB of extra SI cancellation β€” not for free, but at a calibration cost.

Show Hint

Decompose xDL=WDLsDL\mathbf{x}_{DL} = \mathbf{W}_{DL} \mathbf{s}_{DL} with WDL\mathbf{W}_{DL} a NtΓ—KDLN_t \times K_{DL} matrix.

The SI residue is HSIWDLsDL\mathbf{H}_{SI}\mathbf{W}_{DL}\mathbf{s}_{DL}. Optimize WDL\mathbf{W}_{DL} subject to downlink user SINR constraints.

The available null space has dimension Ntβˆ’KDLN_t - K_{DL}; project the downlink beamformer onto it.

Proof
Null-space projection

With KDLK_{DL} active downlink streams and NtN_t Tx antennas, the downlink precoder WDL\mathbf{W}_{DL} lives in a NtΓ—KDLN_t \times K_{DL} subspace of CNt\mathbb{C}^{N_t}. The Ntβˆ’KDLN_t - K_{DL} remaining DoF can be allocated to align with the null space of HSI\mathbf{H}_{SI}.

SI power bound

The residual SI after null-projection under perfect calibration is zero. Calibration error Ο΅cal\epsilon_{\text{cal}} produces βˆ₯Ξ”HSIβˆ₯F2=Ο΅calβˆ₯HSIβˆ₯F2\|\Delta \mathbf{H}_{SI}\|_F^2 = \epsilon_{\text{cal}} \|\mathbf{H}_{SI}\|_F^2, and the residual SI power scales as ΞΊSI2Ο΅calPt\kappa_{\text{SI}}^2 \epsilon_{\text{cal}} P_t with the inverse of the free DoF count.

Conclusion

Combining, the residual SI power is bounded by PtΞΊSI2Ο΅cal/(Ntβˆ’KDL)P_t \kappa_{\text{SI}}^2 \epsilon_{\text{cal}}/(N_t - K_{DL}). For Nt=128,KDL=8N_t = 128, K_{DL} = 8, that is a 10log⁑10(120)β‰ˆ20.810\log_{10}(120) \approx 20.8 dB improvement over a single-antenna baseline β€” adding to the analog/digital cancellation budget without additional hardware. β– \blacksquare

,

Example: Does the Arithmetic Close?

A FD massive MIMO BS has Nt=128N_t = 128 Tx antennas and Nr=32N_r = 32 Rx antennas, operates at f0=3.5f_0 = 3.5 GHz, per-Tx-antenna power Pt/Nt=10P_t/N_t = 10 dBm (so Pttotal=31{P_t}_{\text{total}} = 31 dBm), and the noise floor is βˆ’94-94 dBm over 100 MHz. Assume passive isolation ΞΎ1=45\xi_1 = 45 dB, analog ΞΎ2=35\xi_2 = 35 dB, digital ΞΎ3=35\xi_3 = 35 dB, and spatial nulling from Theorem 27.3.2 gives an additional ΞΎ4=20\xi_4 = 20 dB under moderate calibration error. Does FD close against the noise floor?

Solution
Initial SI power

Per-Rx-antenna unmanaged SI: Pttotalβˆ’ΞΎ1=31βˆ’45=βˆ’14{P_t}_{\text{total}} - \xi_1 = 31 - 45 = -14 dBm. This would saturate any realistic ADC, so analog cancellation is non-negotiable.

After analog stage

βˆ’14βˆ’35=βˆ’49-14 - 35 = -49 dBm. ADC can now digitize cleanly.

After digital stage

βˆ’49βˆ’35=βˆ’84-49 - 35 = -84 dBm. Still 1010 dB above the noise floor βˆ’94-94 dBm β€” unacceptable for reliable uplink reception.

After spatial nulling

βˆ’84βˆ’20=βˆ’104-84 - 20 = -104 dBm, which is 1010 dB below the noise floor. The arithmetic closes, with 1010 dB of margin for imperfections. Total cancellation: ΞΎ=45+35+35+20=135\xi = 45 + 35 + 35 + 20 = 135 dB β€” at the edge of what has been demonstrated in laboratory prototypes.

Interpretation

The budget is theoretically achievable, but every stage assumed an idealization. Real-world phase noise, calibration drift, nonlinear power amplifier distortion, and array coupling all eat into the 10 dB margin. Commercial FD massive MIMO is two to three engineering cycles away from being practical, but nothing in the physics says it is impossible. β– \blacksquare

Common Mistake: Phase Noise Is the Hidden Bottleneck

Mistake:

A common mistake is to treat the digital cancellation stage as software-only and assume arbitrary cancellation depth by increasing computational effort.

Correction:

Digital cancellation is bounded by the phase noise floor of the Tx and Rx local oscillators. Each local oscillator introduces a phase noise PSD of typically βˆ’110-110 dBc/Hz at 100100 kHz offset. Even a perfect algorithmic cancellation leaves a residual SI proportional to the integrated phase noise power, which for typical COTS LO chains caps ΞΎ3\xi_3 at about 5050 dB regardless of algorithm. Overcoming this requires higher-quality (more expensive) LOs, or common-LO architectures that correlate Tx and Rx phase noise so it cancels.

Common Mistake: Power Amplifier Nonlinearity Breaks Linear Cancellation

Mistake:

Treating HSI\mathbf{H}_{SI} as a linear time-invariant filter lets you claim perfect cancellation via adaptive filters. This is the claim in most idealized papers.

Correction:

The Tx power amplifier operates at 22-33 dB compression for efficiency, producing third-order and fifth-order intermodulation products that are not captured by any linear SI channel model. Digital predistortion mitigates but does not eliminate them; residual PA nonlinearity typically leaves 55-1010 dB of SI that linear cancellation cannot touch. The ADS-R and Wits papers since 2020 explicitly model this term; any credible FD budget analysis must include it.

⚠️Engineering Note

Why We Still Don't Have Commercial Full-Duplex Massive MIMO

Full-duplex wireless was demonstrated at Rice (2010), Stanford (2012), and in massive MIMO form at University of Texas-Austin and Lund (2018-2021). Despite 15 years of laboratory progress, no commercial 5G base station supports FD massive MIMO. The reasons are an honest engineering cost-benefit analysis:

  1. Calibration cost: maintaining Ο΅cal≀10βˆ’4\epsilon_{\text{cal}} \leq 10^{-4} across 128128 antennas requires per-antenna calibration every ∼100\sim 100 ms, eating 55-1010 percent of air-interface time.
  2. Noise figure: analog cancellation circuitry adds β‰ˆ2\approx 2 dB to the receiver noise figure, eroding part of the spectral efficiency doubling.
  3. Cross-link interference: FD doubles the exposure to other BS's downlink and other UE's uplink. Without network-wide duplex coordination, the doubled SE evaporates.
  4. Cost: FD requires one RF chain per antenna (no time-sharing), versus TDD massive MIMO which can share chains.

The research agenda for the next decade is to bring all four numbers to acceptable commercial values simultaneously. That is a research program, not a single paper.

Practical Constraints
  • β€’

    Per-antenna calibration dwell: target ≀1\leq 1 ms per refresh

  • β€’

    Receiver noise figure: target degradation ≀0.5\leq 0.5 dB from FD circuitry

  • β€’

    Cross-link interference: requires network-wide duplex coordination (not yet standardized)

  • β€’

    Spatial null calibration: tolerance budget Ο΅cal≀10βˆ’4\epsilon_{\text{cal}} \leq 10^{-4}

πŸ“‹ Ref: 3GPP TR 38.858 (Study on Full-Duplex Systems), Release 18

Historical Note: The Half-Duplex Assumption That Outlived Its Era

1920-present

Shannon's 1948 paper does not require half-duplex operation β€” his channel model is abstractly memoryless and the duplex constraint is nowhere in the theorems. Yet every practical radio from 1920 to 2010 was half-duplex, because the transmitter drowned the receiver. That constraint was so deeply internalized that entire subfields built on its scarcity: frequency-division duplex standards, time alternation protocols, and the whole apparatus of collision avoidance in 802.11802.11.

Rice University's Shannon-compliant full-duplex demonstration (Choi, Jain et al., 2010) broke the assumption for the first time in a real radio. The 1515 years since have steadily closed the gap between demonstration and product. The irony is that the information theory was never the obstacle β€” the antennas, ADCs, and oscillators were.

The Open Problem, Stated Precisely

Design a full-duplex massive MIMO base station that delivers at least 1.8Γ—1.8\times the half-duplex spectral efficiency (after accounting for calibration overhead, cross-link interference, and noise figure penalty) within the power and cost envelope of a current 5G AAU. The following subproblems are individually tractable and collectively necessary:

  1. Joint antenna-and-analog architecture achieving ΞΎ1+ΞΎ2β‰₯90\xi_1 + \xi_2 \geq 90 dB with a 128128-element array (bandwidth β‰₯100\geq 100 MHz).
  2. Digital cancellation algorithm robust to 55-1010 dB of PA nonlinearity with bounded compute cost.
  3. Spatial null allocation strategy that gracefully trades off downlink SINR for SI suppression at ≀5\leq 5 percent rate loss.
  4. Network-level duplex coordination protocol to control cross-link interference (requires 3GPP standardization).
  5. Per-antenna calibration protocol with ≀1\leq 1 ms dwell and Ο΅cal≀10βˆ’4\epsilon_{\text{cal}} \leq 10^{-4}.
,

Why This Matters: Bridge to Integrated Sensing and Communications

Self-interference cancellation is not a problem only full-duplex communication faces. Monostatic integrated sensing and communications (ISAC, Chapter 24) has the same hardware problem with a different motivation: the transmitter emits the probe signal while the co-located receiver listens for its echo. Every cancellation technique that full-duplex develops transfers directly to monostatic ISAC, and vice versa. Full-duplex may reach production first through the ISAC door: the first commercial radio that cancels its own signal by 130130 dB will do so because it wants to see a radar target, not to double the data rate.

Self-Interference Cancellation Cascade

The three-stage sequence of cancellation techniques (passive isolation, analog/RF cancellation, digital baseband cancellation) applied to residues of the previous stage. Typical total budgets in prototypes: ∼120\sim 120-130130 dB. Commercial FD massive MIMO requires roughly 140140-150150 dB including spatial nulling.

Related: Full Duplex, Beamforming Null, Phase Noise Is the Hidden Bottleneck

Quick Check

A full-duplex massive MIMO BS has Nt=256N_t = 256 antennas and serves KDL=16K_{DL} = 16 downlink users. Approximately how much SI cancellation can the spatial null provide (ignoring calibration errors)?

β‰ˆ0\approx 0 dB β€” the null space is too narrow

β‰ˆ20\approx 20 dB

β‰ˆ24\approx 24 dB

β‰ˆ48\approx 48 dB β€” one 10log⁑1010\log_{10} per antenna

Correction:
β‰ˆ24\approx 24 dB

The available null dimension is Ntβˆ’KDL=240N_t - K_{DL} = 240, and the spatial cancellation is approximately 10log⁑10(240)β‰ˆ23.810\log_{10}(240) \approx 23.8 dB. This adds to whatever passive/analog/digital cancellation the cascade provides.

Scalable Distributed Processing for Ultra-Dense Cell-FreeHolographic MIMO Surfaces